U.S. patent number 11,290,477 [Application Number 16/894,332] was granted by the patent office on 2022-03-29 for hierarchical models using self organizing learning topologies.
This patent grant is currently assigned to Cisco Technology, Inc.. The grantee listed for this patent is Cisco Technology, Inc.. Invention is credited to Gregory Mermoud, Laurent Sartran, Pierre-Andre Savalle, Jean-Philippe Vasseur.
United States Patent |
11,290,477 |
Savalle , et al. |
March 29, 2022 |
Hierarchical models using self organizing learning topologies
Abstract
In one embodiment, a device obtains characteristics of a first
anomaly detection model executed by a first distributed learning
agent in a network. The device receives a query from a second
distributed learning agent in the network that requests
identification of a similar anomaly detection to that of a second
anomaly detection model executed by the second distributed learning
agent. The device identifies, after receiving the query from the
second distributed learning agent, the first anomaly detection
model as being similar to that of the second anomaly detection
model, based on the characteristics of the first anomaly detection
model. The device causes the first anomaly detection model to be
sent to the second distributed learning agent for execution.
Inventors: |
Savalle; Pierre-Andre
(Rueil-Malmaison, FR), Mermoud; Gregory (Veyras,
CH), Sartran; Laurent (Palaiseau, FR),
Vasseur; Jean-Philippe (Saint Martin D'uriage, FR) |
Applicant: |
Name |
City |
State |
Country |
Type |
Cisco Technology, Inc. |
San Jose |
CA |
US |
|
|
Assignee: |
Cisco Technology, Inc. (San
Jose, CA)
|
Family
ID: |
58489479 |
Appl.
No.: |
16/894,332 |
Filed: |
June 5, 2020 |
Prior Publication Data
|
|
|
|
Document
Identifier |
Publication Date |
|
US 20200304530 A1 |
Sep 24, 2020 |
|
Related U.S. Patent Documents
|
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
Issue Date |
|
|
16190756 |
Nov 14, 2018 |
10701095 |
|
|
|
15176652 |
Dec 25, 2018 |
10164991 |
|
|
|
62313322 |
Mar 25, 2016 |
|
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
H04L
63/1458 (20130101); H04L 63/1425 (20130101); H04L
63/1416 (20130101); H04L 41/142 (20130101); H04L
63/0236 (20130101) |
Current International
Class: |
H04L
41/142 (20220101) |
References Cited
[Referenced By]
U.S. Patent Documents
Other References
European Search Report dated Jul. 12, 2017 in connection with EP
Application No. 17162413. cited by applicant .
VC Dimension; https://en/wikipedia.org/wiki/VC_dimension; Apr. 29,
2016, pp. 1-3. cited by applicant.
|
Primary Examiner: Pearson; David J
Attorney, Agent or Firm: Behmke Innovation Group Heywood;
Kenneth J. Western; Jonathon P.
Parent Case Text
RELATED APPLICATIONS
This application is a Continuation Application of U.S. patent
application Ser. No. 16/190,756, filed Nov. 14, 2018, entitled
"HIERARCHICAL MODELS USING SELF ORGANIZING LEARNING TOPOLOGIES," by
Savalle et al., which claims priority to U.S. patent application
Ser. No. 15/176,652, now issued as U.S. Pat. No. 10,164,991, filed
Jun. 8, 2016, entitled "HIERARCHICAL MODELS USING SELF ORGANIZING
LEARNING TOPOLOGIES," by Savalle et al., which claims priority to
U.S. Provisional Application No. 62/313,322, filed Mar. 25, 2016,
entitled "HIERARCHICAL MODELS USING SELF ORGANIZING LEARNING
TOPOLOGIES," by Savalle et al., the contents of which are hereby
incorporated by reference.
Claims
What is claimed is:
1. A method comprising: obtaining, by a device, characteristics of
a first anomaly detection model executed by a first distributed
learning agent in a network, wherein the characteristics of the
first anomaly detection model are indicative of network traffic
data associated with the first distributed learning agent;
receiving, at the device, a query from a second distributed
learning agent in the network that requests identification of
another distributed learning agent in the network that is similar
to the second distributed learning agent, wherein the second
distributed learning agent executes a second anomaly detection
model; determining, by the device and after receiving the query
from the second distributed learning agent, that the first
distributed learning agent is similar to the second distributed
learning agent, based on a determination that the network traffic
data associated with the first distributed learning agent is
similar to network traffic data associated with the second
distributed learning agent; and in response to determining that the
first distributed learning agent is similar to the second
distributed learning agent, causing, by the device, the first
anomaly detection model to be sent to the second distributed
learning agent for execution.
2. The method as in claim 1, wherein the characteristics of the
first anomaly detection model are indicative of a volume of network
traffic encountered by the first distributed learning agent.
3. The method as in claim 1, wherein the characteristics of the
first anomaly detection model are indicative of applications or
protocols associated with network traffic encountered by the first
distributed learning agent.
4. The method as in claim 1, wherein the characteristics of the
first anomaly detection model comprise an accuracy measure or a
confidence measure for the first anomaly detection model.
5. The method as in claim 1, wherein causing the first anomaly
detection model to be sent to the second distributed learning agent
for execution comprises: receiving, at the device, a request for
the first anomaly detection model; and sending, by the device, the
first anomaly detection model to the second distributed learning
agent.
6. The method as in claim 1, wherein causing the first anomaly
detection model to be sent to the second distributed learning agent
for execution comprises: sending, by the device, a reply to the
second distributed learning agent indicative of the first anomaly
detection model being similar to the second anomaly detection
model, wherein the second distributed learning agent sends a
request for the first anomaly detection model in response to the
reply.
7. The method as in claim 1, wherein the device is a supervisory
and control agent that oversees operations of the first distributed
learning agent and of the second distributed learning agent.
8. The method as in claim 1, wherein the device is the first
distributed learning agent and comprises a router or switch.
9. The method as in claim 1, wherein causing the first anomaly
detection model to be sent to the second distributed learning agent
for execution comprises: computing a cost function associated with
sending the first anomaly detection model to the second distributed
learning agent for execution based on its similarity to the second
anomaly detection model and on a bandwidth impact of sending the
first anomaly detection model to the second distributed learning
agent.
10. The method as in claim 1, further comprising: using the
characteristics of the first anomaly detection model to train a
statistical clustering model, wherein the first anomaly detection
model is identified as being similar to that of the second anomaly
detection model using the statistical clustering model.
11. An apparatus, comprising: one or more network interfaces to
communicate with a network; a processor coupled to the one or more
network interfaces and configured to execute one or more processes;
and a memory configured to store a process that is executable by
the processor, the process when executed configured to: obtain
characteristics of a first anomaly detection model executed by a
first distributed learning agent in a network, wherein the
characteristics of the first anomaly detection model are indicative
of network traffic data associated with the first distributed
learning agent; receive a query from a second distributed learning
agent in the network that requests identification of another
distributed learning agent in the network that is similar to the
second distributed learning agent, wherein the second distributed
learning agent executes a second anomaly detection model;
determine, after receiving the query from the second distributed
learning agent, that the first distributed learning agent is
similar to the second distributed learning agent, based on a
determination that the network traffic data associated with the
first distributed learning agent is similar to network traffic data
associated with the second distributed learning agent; and in
response to determining that the first distributed learning agent
is similar to the second distributed learning agent, cause the
first anomaly detection model to be sent to the second distributed
learning agent for execution.
12. The apparatus as in claim 11, wherein the characteristics of
the first anomaly detection model are indicative of a volume of
network traffic encountered by the first distributed learning
agent.
13. The apparatus as in claim 11, wherein the characteristics of
the first anomaly detection model are indicative of applications or
protocols associated with network traffic encountered by the first
distributed learning agent.
14. The apparatus as in claim 11, wherein the characteristics of
the first anomaly detection model comprise an accuracy measure or a
confidence measure for the first anomaly detection model.
15. The apparatus as in claim 11, wherein the apparatus causes the
first anomaly detection model to be sent to the second distributed
learning agent for execution by: receiving a request for the first
anomaly detection model; and sending the first anomaly detection
model to the second distributed learning agent.
16. The apparatus as in claim 11, wherein the apparatus causes the
first anomaly detection model to be sent to the second distributed
learning agent for execution by: sending a reply to the second
distributed learning agent indicative of the first anomaly
detection model being similar to the second anomaly detection
model, wherein the second distributed learning agent sends a
request for the first anomaly detection model in response to the
reply.
17. The apparatus as in claim 11, wherein the apparatus is a
supervisory and control agent that oversees operations of the first
distributed learning agent and of the second distributed learning
agent.
18. The apparatus as in claim 11, wherein the apparatus is the
first distributed learning agent and comprises a router or
switch.
19. The apparatus as in claim 11, wherein causing the first anomaly
detection model to be sent to the second distributed learning agent
for execution comprises: computing a cost function associated with
sending the first anomaly detection model to the second distributed
learning agent for execution based on its similarity to the second
anomaly detection model and on a bandwidth impact of sending the
first anomaly detection model to the second distributed learning
agent.
20. A tangible, non-transitory, computer-readable medium storing
program instructions that cause a device in a network to execute a
process comprising: obtaining, by the device, characteristics of a
first anomaly detection model executed by a first distributed
learning agent in a network, wherein the characteristics of the
first anomaly detection model are indicative of network traffic
data associated with the first distributed learning agent;
receiving, at the device, a query from a second distributed
learning agent in the network that requests identification of
another distributed learning agent in the network that is similar
to the second distributed learning agent, wherein the second
distributed learning agent executes a second anomaly detection
model; determining, by the device and after receiving the query
from the second distributed learning agent, that the first
distributed leaning agent is similar to the second distributed
leaning agent, based on a determination that the network traffic
data associated with the first distributed learning agent is
similar to network traffic data associated with the second
distributed learning agent; and in response to determining that the
first distributed learning agent is similar to the second
distributed learning agent, causing, by the device, the first
anomaly detection model to be sent to the second distributed
learning agent for execution.
Description
TECHNICAL FIELD
The present disclosure relates generally to computer networks, and,
more particularly, to hierarchical models organizing learning
topologies in a computer network.
BACKGROUND
Generally, Internet Behavioral Analytics (IBA) refers to the use of
advanced analytics coupled with various networking technologies, to
detect anomalies in a network. Such anomalies may include, for
example, network attacks, malware, misbehaving and misconfigured
devices, and the like. For example, the ability to model the
behavior of a device (e.g., a host, networking switch, router,
etc.) allows for the detection of malware, which is complimentary
to the use of firewalls that use static signature. Observing
behavioral changes (e.g., deviation from modeled behavior) using
flows records, deep packet inspection, and the like, lows for the
detection of an anomaly such as an horizontal movement (e.g.
propagation of a malware, . . . ) or an attempt to perform
information exfiltration, prompting the system to take remediation
actions automatically.
One type of network attack that is of particular concern in the
context of computer networks is a Denial of Service (DoS) attack.
In general, the goal of a DoS attack is to prevent legitimate use
of the services available on the network. For example, a DoS
jamming attack may artificially introduce interference into the
network, thereby causing collisions with legitimate traffic and
preventing message decoding. In another example, a DoS attack may
attempt to overwhelm the network's resources by flooding the
network with requests, to prevent legitimate requests from being
processed. A DoS attack may also be distributed, to conceal the
presence of the attack. For example, a distributed DoS (DDoS)
attack may involve multiple attackers sending malicious requests,
making it more difficult to distinguish when an attack is underway.
When viewed in isolation, a particular one of such a request may
not appear to be malicious. However, in the aggregate, the requests
may overload a resource, thereby impacting legitimate requests sent
to the resource.
Botnets represent one way in which a DDoS attack may be launched
against a network. In a botnet, a subset of the network devices may
be infected with malicious software, thereby allowing the devices
in the botnet to be controlled by a single master. Using this
control, the master can then coordinate the attack against a given
network resource.
BRIEF DESCRIPTION OF THE DRAWINGS
The embodiments herein may be better understood by referring to the
following description in conjunction with the accompanying drawings
in which like reference numerals indicate identically or
functionally similar elements, of which:
FIGS. 1A-1B illustrate an example communication network;
FIG. 2 illustrates an example network device/node;
FIG. 3 illustrates an example self learning network (SLN)
infrastructure;
FIG. 4 illustrates an example distributed learning agent (DLA);
FIGS. 5A-5E illustrate an example of a DLA dynamically swapping
anomaly detection models;
FIGS. 6A-6D illustrate an example of a DLA using model competition
bucket groups;
FIGS. 7A-7C illustrate examples of anomaly score reports;
FIGS. 8A-8B illustrate examples of a DLA sharing anomaly score
distributions;
FIGS. 9A-9D illustrate examples of DLAs sharing anomaly detection
models; and
FIG. 10 illustrates an example simplified procedure for using
anomaly detection models.
DESCRIPTION OF EXAMPLE EMBODIMENTS
Overview
According to one or more embodiments of the disclosure, a device
obtains characteristics of a first anomaly detection model executed
by a first distributed learning agent in a network. The device
receives a query from a second distributed learning agent in the
network that requests identification of a similar anomaly detection
to that of a second anomaly detection model executed by the second
distributed learning agent. The device identifies, after receiving
the query from the second distributed learning agent, the first
anomaly detection model as being similar to that of the second
anomaly detection model, based on the characteristics of the first
anomaly detection model. The device causes the first anomaly
detection model to be sent to the second distributed learning agent
for execution.
DESCRIPTION
A computer network is a geographically distributed collection of
nodes interconnected by communication links and segments for
transporting data between end nodes, such as personal computers and
workstations, or other devices, such as sensors, etc. Many types of
networks are available, with the types ranging from local area
networks (LANs) to wide area networks (WANs). LANs typically
connect the nodes over dedicated private communications links
located in the same general physical location, such as a building
or campus. WANs, on the other hand, typically connect
geographically dispersed nodes over long-distance communications
links, such as common carrier telephone lines, optical lightpaths,
synchronous optical networks (SONET), or synchronous digital
hierarchy (SDH) links, or Powerline Communications (PLC) such as
IEEE 61334, IEEE P1901.2, and others. The Internet is an example of
a WAN that connects disparate networks throughout the world,
providing global communication between nodes on various networks.
The nodes typically communicate over the network by exchanging
discrete frames or packets of data according to predefined
protocols, such as the Transmission Control Protocol/Internet
Protocol (TCP/IP). In this context, a protocol consists of a set of
rules defining how the nodes interact with each other. Computer
networks may be further interconnected by an intermediate network
node, such as a router, to extend the effective "size" of each
network.
Smart object networks, such as sensor networks, in particular, are
a specific type of network having spatially distributed autonomous
devices such as sensors, actuators, etc., that cooperatively
monitor physical or environmental conditions at different
locations, such as, e.g., energy/power consumption, resource
consumption (e.g., water/gas/etc. for advanced metering
infrastructure or "AMI" applications) temperature, pressure,
vibration, sound, radiation, motion, pollutants, etc. Other types
of smart objects include actuators, e.g., responsible for turning
on/off an engine or perform any other actions. Sensor networks, a
type of smart object network, are typically shared-media networks,
such as wireless or PLC networks. That is, in addition to one or
more sensors, each sensor device (node) in a sensor network may
generally be equipped with a radio transceiver or other
communication port such as PLC, a microcontroller, and an energy
source, such as a battery. Often, smart object networks are
considered field area networks (FANs), neighborhood area networks
(NANs), personal area networks (PANs), etc. Generally, size and
cost constraints on smart object nodes (e.g., sensors) result in
corresponding constraints on resources such as energy, memory,
computational speed and bandwidth.
FIG. 1A is a schematic block diagram of an example computer network
100 illustratively comprising nodes/devices, such as a plurality of
routers/devices interconnected by links or networks, as shown. For
example, customer edge (CE) routers 110 may be interconnected with
provider edge (PE) routers 120 (e.g., PE-1, PE-2, and PE-3) in
order to communicate across a core network, such as an illustrative
network backbone 130. For example, routers 110, 120 may be
interconnected by the public Internet, a multiprotocol label
switching (MPLS) virtual private network (VPN), or the like. Data
packets 140 (e.g., traffic/messages) may be exchanged among the
nodes/devices of the computer network 100 over links using
predefined network communication protocols such as the Transmission
Control Protocol/Internet Protocol (TCP/IP), User Datagram Protocol
(UDP), Asynchronous Transfer Mode (ATM) protocol, Frame Relay
protocol, or any other suitable protocol. Those skilled in the art
will understand that any number of nodes, devices, links, etc. may
be used in the computer network, and that the view shown herein is
for simplicity.
In some implementations, a router or a set of routers may be
connected to a private network (e.g., dedicated leased lines, an
optical network, etc.) or a virtual private network (VPN), such as
an MPLS VPN thanks to a carrier network, via one or more links
exhibiting very different network and service level agreement
characteristics. For the sake of illustration, a given customer
site may fall under any of the following categories:
1.) Site Type A: a site connected to the network (e.g., via a
private or VPN link) using a single CE router and a single link,
with potentially a backup link (e.g., a 3G/4G/LTE backup
connection). For example, a particular CE router 110 shown in
network 100 may support a given customer site, potentially also
with a backup link, such as a wireless connection.
2.) Site Type B: a site connected to the network using two MPLS VPN
links (e.g., from different Service Providers), with potentially a
backup link (e.g., a 3G/4G/LTE connection). A site of type B may
itself be of different types:
2a.) Site Type B 1: a site connected to the network using two MPLS
VPN links (e.g., from different Service Providers), with
potentially a backup link (e.g., a 3G/4G/LTE connection).
2b.) Site Type B2: a site connected to the network using one MPLS
VPN link and one link connected to the public Internet, with
potentially a backup link (e.g., a 3G/4G/LTE connection). For
example, a particular customer site may be connected to network 100
via PE-3 and via a separate Internet connection, potentially also
with a wireless backup link.
2c.) Site Type B3: a site connected to the network using two links
connected to the public Internet, with potentially a backup link
(e.g., a 3G/4G/LTE connection).
Notably, MPLS VPN links are usually tied to a committed service
level agreement, whereas Internet links may either have no service
level agreement at all or a loose service level agreement (e.g., a
"Gold Package" Internet service connection that guarantees a
certain level of performance to a customer site).
3.) Site Type C: a site of type B (e.g., types B1, B2 or B3) but
with more than one CE router (e.g., a first CE router connected to
one link while a second CE router is connected to the other link),
and potentially a backup link (e.g., a wireless 3G/4G/LTE backup
link). For example, a particular customer site may include a first
CE router 110 connected to PE-2 and a second CE router 110
connected to PE-3.
FIG. 1B illustrates an example of network 100 in greater detail,
according to various embodiments. As shown, network backbone 130
may provide connectivity between devices located in different
geographical areas and/or different types of local networks. For
example, network 100 may comprise local/branch networks 160, 162
that include devices/nodes 10-16 and devices/nodes 18-20,
respectively, as well as a data center/cloud environment 150 that
includes servers 152-154. Notably, local networks 160-162 and data
center/cloud environment 150 may be located in different geographic
locations.
Servers 152-154 may include, in various embodiments, a network
management server (NMS), a dynamic host configuration protocol
(DHCP) server, a constrained application protocol (CoAP) server, an
outage management system (OMS), an application policy
infrastructure controller (APIC), an application server, etc. As
would be appreciated, network 100 may include any number of local
networks, data centers, cloud environments, devices/nodes, servers,
etc.
In some embodiments, the techniques herein may be applied to other
network topologies and configurations. For example, the techniques
herein may be applied to peering points with high-speed links, data
centers, etc.
In various embodiments, network 100 may include one or more mesh
networks, such as an Internet of Things network. Loosely, the term
"Internet of Things" or "IoT" refers to uniquely identifiable
objects (things) and their virtual representations in a
network-based architecture. In particular, the next frontier in the
evolution of the Internet is the ability to connect more than just
computers and communications devices, but rather the ability to
connect "objects" in general, such as lights, appliances, vehicles,
heating, ventilating, and air-conditioning (HVAC), windows and
window shades and blinds, doors, locks, etc. The "Internet of
Things" thus generally refers to the interconnection of objects
(e.g., smart objects), such as sensors and actuators, over a
computer network (e.g., via IP), which may be the public Internet
or a private network.
Notably, shared-media mesh networks, such as wireless or PLC
networks, etc., are often on what is referred to as Low-Power and
Lossy Networks (LLNs), which are a class of network in which both
the routers and their interconnect are constrained: LLN routers
typically operate with constraints, e.g., processing power, memory,
and/or energy (battery), and their interconnects are characterized
by, illustratively, high loss rates, low data rates, and/or
instability. LLNs are comprised of anything from a few dozen to
thousands or even millions of LLN routers, and support
point-to-point traffic (between devices inside the LLN),
point-to-multipoint traffic (from a central control point such at
the root node to a subset of devices inside the LLN), and
multipoint-to-point traffic (from devices inside the LLN towards a
central control point). Often, an IoT network is implemented with
an LLN-like architecture. For example, as shown, local network 160
may be an LLN in which CE-2 operates as a root node for
nodes/devices 10-16 in the local mesh, in some embodiments.
In contrast to traditional networks, LLNs face a number of
communication challenges. First, LLNs communicate over a physical
medium that is strongly affected by environmental conditions that
change over time. Some examples include temporal changes in
interference (e.g., other wireless networks or electrical
appliances), physical obstructions (e.g., doors opening/closing,
seasonal changes such as the foliage density of trees, etc.), and
propagation characteristics of the physical media (e.g.,
temperature or humidity changes, etc.). The time scales of such
temporal changes can range between milliseconds (e.g.,
transmissions from other transceivers) to months (e.g., seasonal
changes of an outdoor environment). In addition, LLN devices
typically use low-cost and low-power designs that limit the
capabilities of their transceivers. In particular, LLN transceivers
typically provide low throughput. Furthermore, LLN transceivers
typically support limited link margin, making the effects of
interference and environmental changes visible to link and network
protocols. The high number of nodes in LLNs in comparison to
traditional networks also makes routing, quality of service (QoS),
security, network management, and traffic engineering extremely
challenging, to mention a few.
FIG. 2 is a schematic block diagram of an example node/device 200
that may be used with one or more embodiments described herein,
e.g., as any of the computing devices shown in FIGS. 1A-1B,
particularly the PE routers 120, CE routers 110, nodes/device
10-20, servers 152-154 (e.g., a network controller located in a
data center, etc.), any other computing device that supports the
operations of network 100 (e.g., switches, etc.), or any of the
other devices referenced below. The device 200 may also be any
other suitable type of device depending upon the type of network
architecture in place, such as IoT nodes, etc. Device 200 comprises
one or more network interfaces 210, one or more processors 220, and
a memory 240 interconnected by a system bus 250, and is powered by
a power supply 260.
The network interfaces 210 include the mechanical, electrical, and
signaling circuitry for communicating data over physical links
coupled to the network 100. The network interfaces may be
configured to transmit and/or receive data using a variety of
different communication protocols. Notably, a physical network
interface 210 may also be used to implement one or more virtual
network interfaces, such as for virtual private network (VPN)
access, known to those skilled in the art.
The memory 240 comprises a plurality of storage locations that are
addressable by the processor(s) 220 and the network interfaces 210
for storing software programs and data structures associated with
the embodiments described herein. The processor 220 may comprise
necessary elements or logic adapted to execute the software
programs and manipulate the data structures 245. An operating
system 242 (e.g., the Internetworking Operating System, or
IOS.RTM., of Cisco Systems, Inc., another operating system, etc.),
portions of which are typically resident in memory 240 and executed
by the processor(s), functionally organizes the node by, inter
alia, invoking network operations in support of software processors
and/or services executing on the device. These software processors
and/or services may comprise routing process 244 (e.g., routing
services) and illustratively, a self learning network (SLN) process
248, as described herein, any of which may alternatively be located
within individual network interfaces.
It will be apparent to those skilled in the art that other
processor and memory types, including various computer-readable
media, may be used to store and execute program instructions
pertaining to the techniques described herein. Also, while the
description illustrates various processes, it is expressly
contemplated that various processes may be embodied as modules
configured to operate in accordance with the techniques herein
(e.g., according to the functionality of a similar process).
Further, while processes may be shown and/or described separately,
those skilled in the art will appreciate that processes may be
routines or modules within other processes.
Routing process/services 244 include computer executable
instructions executed by processor 220 to perform functions
provided by one or more routing protocols, such as the Interior
Gateway Protocol (IGP) (e.g., Open Shortest Path First, "OSPF," and
Intermediate-System-to-Intermediate-System, "IS-IS"), the Border
Gateway Protocol (BGP), etc., as will be understood by those
skilled in the art. These functions may be configured to manage a
forwarding information database including, e.g., data used to make
forwarding decisions. In particular, changes in the network
topology may be communicated among routers 200 using routing
protocols, such as the conventional OSPF and IS-IS link-state
protocols (e.g., to "converge" to an identical view of the network
topology).
Notably, routing process 244 may also perform functions related to
virtual routing protocols, such as maintaining VRF instance, or
tunneling protocols, such as for MPLS, generalized MPLS (GMPLS),
etc., each as will be understood by those skilled in the art. Also,
EVPN, e.g., as described in the IETF Internet Draft entitled "BGP
MPLS Based Ethernet VPN"<draft-ietf-12vpn-evpn>, introduce a
solution for multipoint L2VPN services, with advanced multi-homing
capabilities, using BGP for distributing customer/client media
access control (MAC) address reach-ability information over the
core MPLS/IP network.
SLN process 248 includes computer executable instructions that,
when executed by processor(s) 220, cause device 200 to perform
anomaly detection functions as part of an anomaly detection
infrastructure within the network. In general, anomaly detection
attempts to identify patterns that do not conform to an expected
behavior. For example, in one embodiment, the anomaly detection
infrastructure of the network may be operable to detect network
attacks (e.g., DDoS attacks, the use of malware such as viruses,
rootkits, etc.). However, anomaly detection in the context of
computer networking typically presents a number of challenges: 1.)
a lack of a ground truth (e.g., examples of normal vs. abnormal
network behavior), 2.) being able to define a "normal" region in a
highly dimensional space can be challenging, 3.) the dynamic nature
of the problem due to changing network behaviors/anomalies, 4.)
malicious behaviors such as malware, viruses, rootkits, etc. may
adapt in order to appear "normal," and 5.) differentiating between
noise and relevant anomalies is not necessarily possible from a
statistical standpoint, but typically also requires domain
knowledge.
Anomalies may also take a number of forms in a computer network:
1.) point anomalies (e.g., a specific data point is abnormal
compared to other data points), 2.) contextual anomalies (e.g., a
data point is abnormal in a specific context but not when taken
individually), or 3.) collective anomalies (e.g., a collection of
data points is abnormal with regards to an entire set of data
points). Generally, anomaly detection refers to the ability to
detect an anomaly that could be triggered by the presence of
malware attempting to access data (e.g., data exfiltration),
spyware, ransom-ware, etc. and/or non-malicious anomalies such as
misconfigurations or misbehaving code. Particularly, an anomaly may
be raised in a number of circumstances: Security threats: the
presence of a malware using unknown attacks patterns (e.g., no
static signatures) may lead to modifying the behavior of a host in
terms of traffic patterns, graphs structure, etc. Machine learning
processes may detect these types of anomalies using advanced
approaches capable of modeling subtle changes or correlation
between changes (e.g., unexpected behavior) in a highly dimensional
space. Such anomalies are raised in order to detect, e.g., the
presence of a 0-day malware, malware used to perform data
ex-filtration thanks to a Command and Control (C2) channel, or even
to trigger (Distributed) Denial of Service (DoS) such as DNS
reflection, UDP flood, HTTP recursive get, etc. In the case of a
(D)DoS, although technical an anomaly, the term "DoS" is usually
used. SLN process 248 may detect malware based on the corresponding
impact on traffic, host models, graph-based analysis, etc., when
the malware attempts to connect to a C2 channel, attempts to move
laterally, or exfiltrate information using various techniques.
Misbehaving devices: a device such as a laptop, a server of a
network device (e.g., storage, router, switch, printer, etc.) may
misbehave in a network for a number of reasons: 1.) a user using a
discovery tool that performs (massive) undesirable scanning in the
network (in contrast with a lawful scanning by a network management
tool performing device discovery), 2.) a software defect (e.g. a
switch or router dropping packet because of a corrupted RIB/FIB or
the presence of a persistent loop by a routing protocol hitting a
corner case). Dramatic behavior change: the introduction of a new
networking or end-device configuration, or even the introduction of
a new application may lead to dramatic behavioral changes. Although
technically not anomalous, an SLN-enabled node having computed
behavioral model(s) may raise an anomaly when detecting a brutal
behavior change. Note that in such as case, although an anomaly may
be raised, a learning system such as SLN is expected to learn the
new behavior and dynamically adapts according to potential user
feedback. Misconfigured devices: a configuration change may trigger
an anomaly: a misconfigured access control list (ACL), route
redistribution policy, routing policy, QoS policy maps, or the
like, may have dramatic consequences such a traffic black-hole, QoS
degradation, etc. SLN process 248 may advantageously identify these
forms of misconfigurations, in order to be detected and fixed.
In various embodiments, SLN process 248 may utilize machine
learning techniques, to perform anomaly detection in the network.
In general, machine learning is concerned with the design and the
development of techniques that take as input empirical data (such
as network statistics and performance indicators), and recognize
complex patterns in these data. One very common pattern among
machine learning techniques is the use of an underlying model M,
whose parameters are optimized for minimizing the cost function
associated to M, given the input data. For instance, in the context
of classification, the model M may be a straight line that
separates the data into two classes (e.g., labels) such that
M=a*x+b*y+c and the cost function would be the number of
misclassified points. The learning process then operates by
adjusting the parameters a,b,c such that the number of
misclassified points is minimal. After this optimization phase (or
learning phase), the model M can be used very easily to classify
new data points. Often, M is a statistical model, and the cost
function is inversely proportional to the likelihood of M, given
the input data.
Computational entities that rely on one or more machine learning
techniques to perform a task for which they have not been
explicitly programmed to perform are typically referred to as
learning machines. In particular, learning machines are capable of
adjusting their behavior to their environment. For example, a
learning machine may dynamically make future predictions based on
current or prior network measurements, may make control decisions
based on the effects of prior control commands, etc.
For purposes of anomaly detection in a network, a learning machine
may construct a model of normal network behavior, to detect data
points that deviate from this model. For example, a given model
(e.g., a supervised, un-supervised, or semi-supervised model) may
be used to generate and report anomaly scores to another device.
Example machine learning techniques that may be used to construct
and analyze such a model may include, but are not limited to,
nearest neighbor (NN) techniques (e.g., k-NN models, replicator NN
models, etc.), statistical techniques (e.g., Bayesian networks,
etc.), clustering techniques (e.g., k-means, etc.), neural networks
(e.g., reservoir networks, artificial neural networks, etc.),
support vector machines (SVMs), or the like.
One class of machine learning techniques that is of particular use
in the context of anomaly detection is clustering. Generally
speaking, clustering is a family of techniques that seek to group
data according to some typically predefined notion of similarity.
For instance, clustering is a very popular technique used in
recommender systems for grouping objects that are similar in terms
of people's taste (e.g., because you watched X, you may be
interested in Y, etc.). Typical clustering algorithms are k-means,
density based spatial clustering of applications with noise
(DBSCAN) and mean-shift, where a distance to a cluster is computed
with the hope of reflecting a degree of anomaly (e.g., using a
Euclidian distance and a cluster based local outlier factor that
takes into account the cluster density).
Replicator techniques may also be used for purposes of anomaly
detection. Such techniques generally attempt to replicate an input
in an unsupervised manner by projecting the data into a smaller
space (e.g., compressing the space, thus performing some
dimensionality reduction) and then reconstructing the original
input, with the objective of keeping the "normal" pattern in the
low dimensional space. Example techniques that fall into this
category include principal component analysis (PCA) (e.g., for
linear models), multi-layer perceptron (MLP) ANNs (e.g., for
non-linear models), and replicating reservoir networks (e.g., for
non-linear models, typically for time series).
According to various embodiments, SLN process 248 may also use
graph-based models for purposes of anomaly detection. Generally
speaking, a graph-based model attempts to represent the
relationships between different entities as a graph of nodes
interconnected by edges. For example, ego-centric graphs have been
used to represent the relationship between a particular social
networking profile and the other profiles connected to it (e.g.,
the connected "friends" of a user, etc.). The patterns of these
connections can then be analyzed for purposes of anomaly detection.
For example, in the social networking context, it may be considered
anomalous for the connections of a particular profile not to share
connections, as well. In other words, a person's social connections
are typically also interconnected. If no such interconnections
exist, this may be deemed anomalous.
An example self learning network (SLN) infrastructure that may be
used to detect network anomalies is shown in FIG. 3, according to
various embodiments. Generally, network devices may be configured
to operate as part of an SLN infrastructure to detect, analyze,
and/or mitigate network anomalies such as network attacks (e.g., by
executing SLN process 248). Such an infrastructure may include
certain network devices acting as distributed learning agents
(DLAs) and one or more supervisory/centralized devices acting as a
supervisory and control agent (SCA). A DLA may be operable to
monitor network conditions (e.g., router states, traffic flows,
etc.), perform anomaly detection on the monitored data using one or
more machine learning models, report detected anomalies to the SCA,
and/or perform local mitigation actions. Similarly, an SCA may be
operable to coordinate the deployment and configuration of the DLAs
(e.g., by downloading software upgrades to a DLA, etc.), receive
information from the DLAs (e.g., detected anomalies/attacks,
compressed data for visualization, etc.), provide information
regarding a detected anomaly to a user interface (e.g., by
providing a webpage to a display, etc.), and/or analyze data
regarding a detected anomaly using more CPU intensive machine
learning processes.
One type of network attack that is of particular concern in the
context of computer networks is a Denial of Service (DoS) attack.
In general, the goal of a DoS attack is to prevent legitimate use
of the services available on the network. For example, a DoS
jamming attack may artificially introduce interference into the
network, thereby causing collisions with legitimate traffic and
preventing message decoding. In another example, a DoS attack may
attempt to overwhelm the network's resources by flooding the
network with requests (e.g., SYN flooding, sending an overwhelming
number of requests to an HTTP server, etc.), to prevent legitimate
requests from being processed. A DoS attack may also be
distributed, to conceal the presence of the attack. For example, a
distributed DoS (DDoS) attack may involve multiple attackers
sending malicious requests, making it more difficult to distinguish
when an attack is underway. When viewed in isolation, a particular
one of such a request may not appear to be malicious. However, in
the aggregate, the requests may overload a resource, thereby
impacting legitimate requests sent to the resource.
Botnets represent one way in which a DDoS attack may be launched
against a network. In a botnet, a subset of the network devices may
be infected with malicious software, thereby allowing the devices
in the botnet to be controlled by a single master. Using this
control, the master can then coordinate the attack against a given
network resource.
DoS attacks are relatively easy to detect when they are brute-force
(e.g. volumetric), but, especially when highly distributed, they
may be difficult to distinguish from a flash-crowd (e.g., an
overload of the system due to many legitimate users accessing it at
the same time). This fact, in conjunction with the increasing
complexity of performed attacks, makes the use of "classic"
(usually threshold-based) techniques useless for detecting them.
However, machine learning techniques may still be able to detect
such attacks, before the network or service becomes unavailable.
For example, some machine learning approaches may analyze changes
in the overall statistical behavior of the network traffic (e.g.,
the traffic distribution among flow flattens when a DDoS attack
based on a number of microflows happens). Other approaches may
attempt to statistically characterizing the normal behaviors of
network flows or TCP connections, in order to detect significant
deviations. Classification approaches try to extract features of
network flows and traffic that are characteristic of normal traffic
or malicious traffic, constructing from these features a classifier
that is able to differentiate between the two classes (normal and
malicious).
As shown in FIG. 3, routers CE-2 and CE-3 may be configured as DLAs
and server 152 may be configured as an SCA, in one implementation.
In such a case, routers CE-2 and CE-3 may monitor traffic flows,
router states (e.g., queues, routing tables, etc.), or any other
conditions that may be indicative of an anomaly in network 100. As
would be appreciated, any number of different types of network
devices may be configured as a DLA (e.g., routers, switches,
servers, blades, etc.) or as an SCA.
Assume, for purposes of illustration, that CE-2 acts as a DLA that
monitors traffic flows associated with the devices of local network
160 (e.g., by comparing the monitored conditions to one or more
machine-learning models). For example, assume that device/node 10
sends a particular traffic flow 302 to server 154 (e.g., an
application server, etc.). In such a case, router CE-2 may monitor
the packets of traffic flow 302 and, based on its local anomaly
detection mechanism, determine that traffic flow 302 is anomalous.
Anomalous traffic flows may be incoming, outgoing, or internal to a
local network serviced by a DLA, in various cases.
In some cases, traffic 302 may be associated with a particular
application supported by network 100. Such applications may
include, but are not limited to, automation applications, control
applications, voice applications, video applications,
alert/notification applications (e.g., monitoring applications),
communication applications, and the like. For example, traffic 302
may be email traffic, HTTP traffic, traffic associated with an
enterprise resource planning (ERP) application, etc.
In various embodiments, the anomaly detection mechanisms in network
100 may use Internet Behavioral Analytics (IBA). In general, IBA
refers to the use of advanced analytics coupled with networking
technologies, to detect anomalies in the network. Although
described later with greater details, the ability to model the
behavior of a device (networking switch/router, host, etc.) will
allow for the detection of malware, which is complementary to the
use of a firewall that uses static signatures. Observing behavioral
changes (e.g., a deviation from modeled behavior) thanks to
aggregated flows records, deep packet inspection, etc., may allow
detection of an anomaly such as an horizontal movement (e.g.
propagation of a malware, etc.), or an attempt to perform
information exfiltration.
FIG. 4 illustrates an example distributed learning agent (DLA) 400
in greater detail, according to various embodiments. Generally, a
DLA may comprise a series of modules hosting sophisticated tasks
(e.g., as part of an overall SLN process 248). Generally, DLA 400
may communicate with an SCA (e.g., via one or more northbound APIs
402) and any number of nodes/devices in the portion of the network
associated with DLA 400 (e.g., via APIs 420, etc.).
In some embodiments, DLA 400 may execute a Network Sensing
Component (NSC) 416 that is a passive sensing construct used to
collect a variety of traffic record inputs 426 from monitoring
mechanisms deployed to the network nodes. For example, traffic
record inputs 426 may include Cisco.TM. Netflow records,
application identification information from a Cisco.TM. Network
Based Application Recognition (NBAR) process or another
application-recognition mechanism, administrative information from
an administrative reporting tool (ART), local network state
information service sets, media metrics, or the like.
Furthermore, NSC 416 may be configured to dynamically employ Deep
Packet Inspection (DPI), to enrich the mathematical models computed
by DLA 400, a critical source of information to detect a number of
anomalies. Also of note is that accessing control/data plane data
may be of utmost importance, to detect a number of advanced threats
such as data exfiltration. NSC 416 may be configured to perform
data analysis and data enhancement (e.g., the addition of valuable
information to the raw data through correlation of different
information sources). Moreover, NSC 416 may compute various
networking based metrics relevant for the Distributed Learning
Component (DLC) 408, such as a large number of statistics, some of
which may not be directly interpretable by a human.
In some embodiments, DLA 400 may also include DLC 408 that may
perform a number of key operations such as any or all of the
following: computation of Self Organizing Learning Topologies
(SOLT), computation of "features" (e.g., feature vectors), advanced
machine learning processes, etc., which DLA 400 may use in
combination to perform a specific set of tasks. In some cases, DLC
408 may include a reinforcement learning (RL) engine 412 that uses
reinforcement learning to detect anomalies or otherwise assess the
operating conditions of the network. Accordingly, RL engine 412 may
maintain and/or use any number of communication models 410 that
model, e.g., various flows of traffic in the network. In further
embodiments, DLC 408 may use any other form of machine learning
techniques, such as those described previously (e.g., supervised or
unsupervised techniques, etc.). For example, in the context of SLN
for security, DLC 408 may perform modeling of traffic and
applications in the area of the network associated with DLA 400.
DLC 408 can then use the resulting models 410 to detect graph-based
and other forms of anomalies (e.g., by comparing the models with
current network characteristics, such as traffic patterns. The SCA
may also send updates 414 to DLC 408 to update model(s) 410 and/or
RL engine 412 (e.g., based on information from other deployed DLAs,
input from a user, etc.).
When present, RL engine 412 may enable a feed-back loop between the
system and the end user, to automatically adapt the system
decisions to the expectations of the user and raise anomalies that
are of interest to the user (e.g., as received via a user interface
of the SCA). In one embodiment, RL engine 412 may receive a signal
from the user in the form of a numerical reward that represents for
example the level of interest of the user related to a previously
raised event. Consequently the agent may adapt its actions (e.g.
search for new anomalies), to maximize its reward over time, thus
adapting the system to the expectations of the user. More
specifically, the user may optionally provide feedback thanks to a
lightweight mechanism (e.g., `like` or `dislike`) via the user
interface.
In some cases, DLA 400 may include a threat intelligence processor
(TIP) 404 that processes anomaly characteristics so as to further
assess the relevancy of the anomaly (e.g. the applications involved
in the anomaly, location, scores/degree of anomaly for a given
model, nature of the flows, or the like). TIP 404 may also generate
or otherwise leverage a machine learning-based model that computes
a relevance index. Such a model may be used across the network to
select/prioritize anomalies according to the relevancies.
DLA 400 may also execute a Predictive Control Module (PCM) 406 that
triggers relevant actions in light of the events detected by DLC
408. In order words, PCM 406 is the decision maker, subject to
policy. For example, PCM 406 may employ rules that control when DLA
400 is to send information to the SCA (e.g., alerts, predictions,
recommended actions, trending data, etc.) and/or modify a network
behavior itself. For example, PCM 406 may determine that a
particular traffic flow should be blocked (e.g., based on the
assessment of the flow by TIP 404 and DLC 408) and an alert sent to
the SCA.
Network Control Component (NCC) 418 is a module configured to
trigger any of the actions determined by PCM 406 in the network
nodes associated with DLA 400. In various embodiments, NCC 418 may
communicate the corresponding instructions 422 to the network nodes
using APIs 420 (e.g., DQoS interfaces, ABR interfaces, DCAC
interfaces, etc.). For example, NCC 418 may send mitigation
instructions 422 to one or more nodes that instruct the receives to
reroute certain anomalous traffic, perform traffic shaping, drop or
otherwise "black hole" the traffic, or take other mitigation steps.
In some embodiments, NCC 418 may also be configured to cause
redirection of the traffic to a "honeypot" device for forensic
analysis. Such actions may be user-controlled, in some cases,
through the use of policy maps and other configurations. Note that
NCC 418 may be accessible via a very flexible interface allowing a
coordinated set of sophisticated actions. In further embodiments,
API(s) 420 of NCC 418 may also gather/receive certain network data
424 from the deployed nodes such as Cisco.TM. OnePK information or
the like.
The various components of DLA 400 may be executed within a
container, in some embodiments, that receives the various data
records and other information directly from the host router or
other networking device. Doing so prevents these records from
consuming additional bandwidth in the external network. This is a
major advantage of such a distributed system over centralized
approaches that require sending large amount of traffic records.
Furthermore, the above mechanisms afford DLA 400 additional insight
into other information such as control plane packet and local
network states that are only available on premise. Note also that
the components shown in FIG. 4 may have a low footprint, both in
terms of memory and CPU. More specifically, DLA 400 may use
lightweight techniques to compute features, identify and classify
observation data, and perform other functions locally without
significantly impacting the functions of the host router or other
networking device.
As noted above, edge devices may perform anomaly detection by
analyzing traffic flows. In the context of building anomaly
detection models at the network edge from network traffic, there is
a central tradeoff between using very specific models and very
general models. Notably, models constructed from all of the traffic
for a given application are very general in nature and can get a
lot of samples if there is enough traffic for the application. As a
consequence, these models tend to be rather confident and accurate
in modeling many different behaviors at the same time. Such models
are essential to detecting very strong network anomalies with high
confidence. However, because they mix up a large number of
behaviors, they tend to miss more subtle anomalies where the
traffic is anomalous for the specific hosts involved.
On the other hand, very specific models such as those based on the
traffic between two hosts, or two groups of hosts, can be useful
for detecting more subtle anomalies that appear only in a specific
context. However, very specific models can also suffer from
low-samples effects, as there may be much less input data to
assess. For example, a very general model may assess all of the
HTTP traffic for the local network, whereas a very specific model
may assess HTTP traffic for only a particular host in the local
network. Thus, the very specific model may be better able to
identify subtle anomalies for the specific host, but also have
significantly smaller set of traffic data to assess. As a
consequence, finding the right scale at which to build statistical
models is difficult.
Hierarchical Models to Organize Learning Topologies
The techniques herein introduce a multi-scale method for network
traffic anomaly detection on distributed learning agents, where
models corresponding to different levels of aggregation and
specificity are built in parallel. In some aspects, all the output
scores of these models are then use to score input events in terms
of how anomalous they are. In addition, the capacity and
dimensionality of machine learning models is adapted dynamically
based on the performance and confidence of the models.
Specifically, according to one or more embodiments of the
disclosure as described in detail below, a device in a network
maintains a plurality of anomaly detection models for different
sets of aggregated traffic data regarding traffic in the network.
The device determines a measure of confidence in a particular one
of the anomaly detection models that evaluates a particular set of
aggregated traffic data. The device dynamically replaces the
particular anomaly detection model with a second anomaly detection
model configured to evaluate the particular set of aggregated
traffic data and has a different model capacity than that of the
particular anomaly detection model. The device provides an anomaly
event notification to a supervisory controller based on a combined
output of the second anomaly detection model and of one or more of
the anomaly detection models in the plurality of anomaly detection
models.
Illustratively, the techniques described herein may be performed by
hardware, software, and/or firmware, such as in accordance with the
SLN process 248, which may include computer executable instructions
executed by the processor 220 (or independent processor of
interfaces 210) to perform functions relating to the techniques
described herein, e.g., in conjunction with routing process
244.
Operationally, a first aspect of the multi-scale mechanism
associates input traffic with model keys, based on a policy. This
component can either process raw network data (e.g., Netflow
records, DPI records, other types of network records, etc.) and/or
aggregated data (e.g., features computed in time bins based on raw
network data). In general, a model key is an identifier of the
model. A given chunk of input data, or bin of aggregated data, can
be associated with multiple model keys. For instance, a Netflow or
other traffic record, or aggregated traffic bin, could be
associated to any or all of the following model keys: Application
classification of traffic Source IP and application classification
of traffic Source MAC address and application classification of
traffic Destination IP and application classification of traffic
Destination MAC address and application classification of traffic
Source IP, destination IP, and application classification of
traffic Source MAC, destination MAC, and application classification
of traffic
Each model key represents a different level of aggregation and mode
construction "scale". In particular, the model keys can be selected
in a non-uniform fashion. For instance, certain levels of
aggregation may be less relevant for some application
classification or hosts involved.
In another embodiment, model keys can also be constructed based on
groups (or, clusters) or hosts provided by an external system (e.g.
a central controller such as an SCA). In this context, IPs and MAC
addresses can be mapped to groups, and additional model keys may be
used, including: Source group and application classification of
traffic Destination group and application classification of traffic
Source and destinations groups, and application classification of
traffic
As indicated above, these multiple scales allow for capturing
different levels of anomalies with different levels of
confidence.
FIGS. 5A-5E illustrate an example of a DLA dynamically swapping
anomaly detection models, in accordance with various embodiments.
In one embodiment model keys may be pre-configured upon set up
whereas in another embodiment model keys may be built up on the fly
upon requests by a central controller using a custom model-key( )
message sent by the central controller to a DLA. For example, as
shown in FIG. 5A, SCA 502 may compute various model keys and
install the model keys to DLA 400 via a model_key( ) message 504.
Such an approach allows for building granular models that are
specific to the criterion such as the DLA location, applications
seen by the DLA, etc.
A second aspect of the multi-scale mechanism involves the
construction of statistical models for each model_key, based on all
the traffic records or aggregated traffic bins corresponding to
these keys. For example, as shown in FIG. 5B, DLA 400 may use
traffic data from the local network to construct anomaly detection
models based on the model keys received from SCA 402. Any type of
statistical model can be used, including approaches based on coding
and reconstruction errors, density estimators, or change point
detection models. In particular, the models may forget about the
past at a rate that does or does not depend on the amount of
traffic.
In one embodiment, the DLA may dynamically tune the modeling
capacity or dimensionality of each statistical model, depending on
the aggregation to which the model corresponds. In machine
learning, capacity refers to the ability of a model to capture
complex behaviors. Although a high capacity model allows to capture
more, it is also much more prone to over fitting, as it can capture
and model patterns that are just random or due to noise. For
instance, it may be desirable to use a higher capacity model for
keys that aggregate a lot of traffic, in order to capture more
behaviors. In particular, model capacities and dimensionalities can
be dynamically adjusted based on the number of samples per model,
the past accuracies of models or other similar metrics. A DLA may
also decide not to build models for a given key if they are
detected as unstable, lacking samples, etc., in which case a custom
message model-status( ) may be used by the DLA to report model
status to the central controller. For example, as shown in FIG. 5C,
DLA 400 may determine that it does not have sufficient traffic data
for a given model key to construct a model. In turn, DLA 400 may
notify SCA 502 that the model cannot be constructed via a
model_status( ) message 506.
As shown in FIG. 5D, DLA 400 may dynamically swap models based on
their capacities and/or accuracies/confidence scores. For example,
when the DLA determines that a given model is considered uncertain
or inaccurate based on the above metrics, the DLA may start up a
lower capacity model corresponding to the same model keys. When the
lower capacity model has been trained, the previous model is
deactivated, in one embodiment. Similarly, when a model has a very
high confidence score, the DLA may start up a higher capacity model
in parallel as a substitute, and the previous model is similarly
retired when the new model is confident enough.
In one embodiment, the size of the region of the models' input
spaces that is considered as anomalous may be estimated using
statistical means, and used to temporarily disable the production
of scores from models where the anomalous region is tiny or empty.
Indeed, some models might degenerate to considering that most
behaviors are normal, as might be the case, and the evaluation of
this model is thus a waste of computational power.
A third aspect of the multi-scale mechanism may monitor the output
scores from the models, and transform them into anomaly events that
can be forwarded to a global anomaly detection system. In
particular, a DLA may combine the outputs of its active models, and
use the combined outputs to report anomaly events. For example, as
shown in FIG. 5E, DLA 400 may send a custom output_report( )
message 508 to SCA 502 that is based on the combination of model
outputs from the active models on DLA 400.
DLA 400 may, based on an internal cost function, decide not to use
scores from a model that is considered uncertain according to some
statistical metric (such as temporal changes in the model, or
overall number of samples used to estimate the model). In addition,
DLA 400 may take into account the amount of samples in the models
to provide a score for the corresponding anomalies. For instance,
input that is anomalous for aggregation levels such as "all the
traffic corresponding to a given aggregation" will be anomalous for
all the model keys that are more specific and can thus be
considered to be a strong anomaly. On the other hand, input that is
anomalous only for a specific pair of hosts or groups of hosts may
be considered anomalous with less confidence, as this stems from
models that are very specific, and may just be a false
positive.
DLA 400 may optionally take into account a variety of external
feed-back such as signals from SCA 502 about qualities of anomalies
for a given key. DLA 400 can store such feedback locally so as to
keep track of the relevancy of anomalies on a per-model/per-key
basis, in order to trigger appropriate action (e.g. cancel a model
that is not efficient enough in order to save CPU and memory
resources). In yet another embodiment, the DLA may leverage an
internal API to take memory and CPU resources into account, to
dynamically control the number of active models according to their
performance and the available local resources.
Distributed Architecture for Extreme Tracking of Anomaly Scores
As noted above, an anomaly detection system may use multiple
statistical or decision models by combining their scores to reach a
final decision. In some cases, the most interesting anomalous
events may only be anomalous for a handful of models or experts.
This is in particular the case for detection of anomalies from
network traffic based on hierarchical models, as described above,
where different models may produce scores on different mutually
exclusive subsets of the input traffic. For instance, there may be
only a subset of models that express scores on a given type of
application
a given application. In such cases, there may only be a handful of
models to express scores on a given chunk of traffic, and requiring
many models to agree is unrealistic. This is unlike the common
wisdom in other applications of such systems, where multiple models
or experts agreeing are the most important.
Additionally, combining many signals and models in the context of
anomaly decision may lead to a lot of anomalies, as considering
more models increases the statistical odds that some of the input
data may be considered anomalous by at least one of them. The
generation of many anomalous events is an issue in most anomaly
detection systems and especially in distributed systems where a
very large number of distributed learning agents may each
contribute anomalies. This is particularly true when the decision
of a single expert must be considered sufficient to deem an input
anomalous. The challenge that stems from these considerations is
that of determining which events are most rare in the context of
many models, which may express scores on different subsets of the
input data.
In particular, the problem is even more acute in a distributed
setting, where the models on each DLA might only have a partial
view of what events are actually rare at the level of the full
system. In addition, memory is usually very limited in distributed
systems, and usually already used for storing statistical models. A
system for combining and regulating the scores of the models should
thus use a very small amount of memory in this context. Similarly,
a distributed sequential anomaly detection system usually has
real-time processing requirements, and should thus be efficient
from a CPU perspective.
In various embodiments, further techniques are introduced that
provide an approach for regulating the production of anomalies from
distributed models, both locally on the DLAs, and globally on a
central controller (e.g., the SCA). More specifically, a system is
introduced for tracking scores from multiple statistical models for
anomaly detection on a DLA, and selecting the most anomalous
events, both from the perspective of the scores seen locally, and
of the scores received from other agents. The system groups the
models by "buckets" where the models compete with each other in
terms of scores produced locally and globally, and model scores
that are low with respect to their assigned buckets are used to
produce anomalies. This process may proceed in an incremental
fashion, by keeping a small state compared to the number of scores
received.
Illustratively, the techniques described herein may be performed by
hardware, software, and/or firmware, such as in accordance with the
SLN process 248, which may include computer executable instructions
executed by the processor 220 (or independent processor of
interfaces 210) to perform functions relating to the techniques
described herein, e.g., in conjunction with routing process
244.
Operationally, the anomaly detection system may track the output
scores of its underlying models, possibly produced on a DLA, and
identify the rarest events, both from the point of view of the
agent and from a global point of view. In some aspects, a first
aspect of the score tracking mechanism may models in "buckets" of
data. This component maintains a relationship between the models,
and a smaller set of buckets, where each bucket corresponds to a
set of models. A model may be assigned to multiple bucket
groups.
FIGS. 6A-6D illustrate an example of a DLA using model competition
bucket groups. The bucketing scheme can be used to create
competition between similar models, according to various
embodiments. For example, very high-level models may be put in one
bucket, while a set of more specific models for a couple very
specific devices and a given protocol may be put in another bucket.
This can be adjusted to the specifics of the anomaly detection
problem considered. As shown in FIG. 6A, DLA 400 may assign its
various anomaly detection models to any number of competition
bucket groups.
In one embodiment, the bucketing can be modified dynamically based
on feedback from a central controller, such as SCA 502. In
particular, a DLA may send summary statistics about the buckets,
including the models currently assigned to a bucket, or the samples
and some summary statistics about the scores contributed by each
model to the bucket. This information may be used by the SCA to
determine if the buckets encode the desired competition structure
between models.
In addition, the first component of the score tracking mechanism
may receive and handle requests to dynamically reshape the buckets.
The following two operations may be supported: (i) merging two
buckets, and/or (ii) splitting a bucket in two new empty buckets.
In the second case, any data assigned to the two buckets to fuse
may either be discarded, or be used to initialize the new buckets.
For this, the devices may send custom unicast/multicast messages
BucketsStatisticsRequest( ), BucketsStatistics( ), and
BucketsReshape( ), to request the summary statistics about the
buckets, to send the summary statistics about the buckets, and to
perform the fuse/split actions on buckets, respectively. For
example, as shown in FIG. 6B, SCA 502 may request bucket
statistics/information from DLA 400 regarding its model buckets
using a BucketsStatisticsRequest( ) message 602. In response, DLA
400 may return statistics regarding its bucket groups via a
BucketStatistics( ) message 604. In addition, as shown in FIG. 6C,
SCA 502 may determine that the buckets of DLA 400 should be
adjusted and send a BucketReshape( ) message 606 to DLA 400 to do
so. As shown in FIG. 6D, DLA 400 may compare intra-bucket models to
assess their output scores.
For purposes of illustration only, examples of the score tracking
mechanism are illustrated in FIGS. 7A-7C. In a first example shown
in illustration 700 of FIG. 7A, consider a single model assigned to
a single bucket. In this special case, there is no competition, and
the bucket mechanism acts primarily as a detector of extreme score
values out of the model. Illustrations 710 shown in FIG. 7B
illustrates the case in which two or more models are assigned to
the same competition bucket group.
Another aspect of the score tracking mechanism may estimate the
most anomalous scores produced over a window by the models in each
bucket, in some embodiments. The scores can be integrated into the
bucket either immediately, or after a predefined delay. In one
embodiment, for each bucket, the device may keep at most K score
instances corresponding to the K most anomalous scores seen in the
bucket during a time window. Usually, anomalous scores are
represented through high score values. This can be achieved through
a simple list of the most anomalous scores, or through a more
efficient data structure that maintains the scores of the bucket
sorted.
The DLA may keep only the scores produced over a given time window
by discarding the scores produced earlier than the time of the
latest score minus the width of the time window. This requires a
careful configuration of K so that the number of scores kept in the
tracker doesn't fall excessively.
FIG. 7A illustrates the operation of the bucket mechanism in the
single model case with no competition and an infinite window. As
shown, the initial contents of the bucket are shown in panel A.
After a score update, the bucket remains unchanged in panel B
because the score is not anomalous enough to be in the top K. In
panel C, a score that is anomalous enough has been observed. The
lowest score (e.g., the least anomalous in the set of most
anomalous scores) is flushed out of the bucket, and the new score
is added to the bucket.
FIG. 7B illustrates the behavior of the bucket approach in the two
model case with competition and an infinite window. The initial
contents of the bucket are shown in panel A. The bucket is then
updated using scores from both models. Only the score from model M1
is anomalous enough to make it to the contents of the bucket. The
same applies to the second update. In this case, model M2 is never
able to generate scores that are anomalous enough to update the
bucket.
In another embodiment, for each bucket, the score tracking
mechanism keeps at most K scores instances corresponding to the
most anomalous scores seen in the bucket during a time window, with
the limitation that only the L most anomalous scores corresponding
to a given key can be kept in the bucket. The function charged with
selecting the scores of interest may either be locally configured
on the DLA or dynamically updated by an SCA with more complex
policies (e.g., keep the K top scores over a period of time of H
hours, allowing for K'>K if all scores are higher than the 99
percentile computed over the past D days, etc.).
FIG. 7C illustrates the behavior of the above bucket mechanism in
the single model case with an infinite window and L=1. For the
illustration, lowercase letters are used as keys. The initial
contents of the bucket are shown in panel A. After a score update,
the bucket remains unchanged in panel B because the score is not
anomalous enough to be in the top K. After another score update,
the bucket still remains unchanged in panel C, because there is
already a more anomalous score for the `a` key. In panel D, a more
anomalous score has been observed for the `a` key, and replaces the
previous one in the bucket. Finally, in panel E, a score
corresponding to a new key has been observed, and was sufficiently
anomalous to be kept in the bucket.
The key is additional information, orthogonal to the model, which
is provided with a score instance. For instance, a network traffic
event may lead to multiple scores instances. In this embodiment, a
key can be used to ensure that the network traffic event only
occupies L slots in the list of scores of the buckets at most. This
can be adjusted to the specifics of the anomaly detection problem
considered.
A third aspect of the score tracking mechanism may send the
contents of the buckets on the DLA along with other information
such as the model keys, etc. to the SCA and/or receive such
information from other DLAs or the SCA. The devices may send most
anomalous score reports for the buckets either periodically or on
request. To this end, the devices may send custom unicast/multicast
messages Buckets_Most_Anomalous_Scores_Distributions_Request( ) and
Buckets_Most_Anomalous_Scores_Distributions( ), to request the
buckets' contents from another entity or send its buckets'
contents, respectively.
FIGS. 8A-8B illustrate examples of a DLA sharing anomaly score
distributions, in various embodiments. As shown in FIG. 8A, DLA
400a may send a Buckets_Most_Anomalous_Scores_Distributions( )
message 802 to SCA 502 and/or any other DLA in the network, such as
DLA 400b, to report its most anomalous scores for its buckets.
Similarly, as shown in FIG. 8B, DLA 400 may send one or more
Buckets_Most_Anomalous_Scores_Distributions_Request( ) messages 804
to SCA 502 or other DLAs, such as DLA 400b, to request their most
anomalous scores.
A fourth aspect of the score tracking mechanism may receive scores
from the models. For each score instance, the ranks of the score in
the buckets where the model is in are computed and aggregated using
an aggregation function such as a minimum threshold. If the
aggregate rank is below a threshold, an anomalous event is
produced. In one embodiment, the ranks are computed only using the
local buckets. In another embodiment, the ranks are computed using
both the local buckets, and any matching buckets that have been
received through the Buckets_Most_Anomalous_Scores_Distributions( )
message.
Dynamic Cooperation of DLAs
Also as noted above, it is paramount that DLAs are able to make
immediate decisions as to whether traffic is suspicious, in order
to gather enough context about the offending traffic, and report
the anomaly without delay. Gathering context may include gathering
raw packet data for the offending traffic, or any other traffic
that is deemed context-relevant by the system. In particular, this
emphasizes that decisions must be taken without much delay directly
by the DLA, and that systematically routing the data through the
SCA or other DLAs is not acceptable. Note that such issues are even
exacerbated in systems such as in the IoT where DLAs are
potentially connected to the controller/SCA via low-speed links
with potentially intermittent connectivity.
A common issue with DLAs is that they only have a partial view of
the network topology and traffic. For anomaly detection based on
machine learning, this can result in statistical models that have
only been trained on a limited amount of data. As a consequence,
these "bird's eye" view models may have lower confidence or
prediction accuracy than models that would have been trained using
more data, possibly from multiple DLAs. This is especially true in
systems that learn and make use of many targeted models, such as in
the hierarchical architecture described above. For instance,
specific models may be built for DNS servers that are observed in
the traffic for groups of desktop machines or for remote cloud
servers with which the local users are interacting. In particular,
a specific model may also be built for a specific business
application, but the corresponding model may only get very limited
samples if the application is not used often by the local
users.
The techniques herein further introduce an approach for dealing
with low-samples regimes and low-confidence models when doing
anomaly detection using machine learning at the network edge. DLAs
may either auto-assess their statistical models or have a
controller (e.g., SCA) assess their statistical models. Such
devices may be configured to identify other DLAs that have
compatible models, to enrich the local model of a given DLA.
Notably, if bandwidth and utility constraints are met, the
compatible models may be requested and combined with the local
model.
Illustratively, the techniques described herein may be performed by
hardware, software, and/or firmware, such as in accordance with the
SLN process 248, which may include computer executable instructions
executed by the processor 220 (or independent processor of
interfaces 210) to perform functions relating to the techniques
described herein, e.g., in conjunction with routing process
244.
Operationally, the cooperation mechanism herein allows DLAs to
exchange models between one another, in order to improve their
real-time prediction accuracy at almost no computational cost. A
first component of the cooperation mechanism may monitor the models
on the DLAs, and assess whether some of them are insufficiently
accurate or confident. For example, as shown in FIG. 9A, SCA 502 or
another device (e.g., a particular DLA, etc.) may receive model
characteristics 902 from DLAs 400 and 400b.
In a first embodiment, a DLA may examine elements such as the
amount of samples, the variance of the predictions, or possibly
feedback from an SCA about previous anomalies signaled by the DLA.
This is referred to herein as an auto-assessment mode.
In another embodiment, an SCA or other remote controller may
perform the above functionality, which may observe the models on
the DLAs. The external controller can take advantage of extra data,
such as from other DLAs, to determine accuracy and confidence
measures for the models in relation to those of other DLAs. This is
referred to herein as the relative assessment mode.
In both embodiments, the assessments may be performed periodically
for a subset of the models maintained at the DLA. The subset can be
selected dynamically based on a cost function depending on the time
since the last assessments, confidence metrics, or
protocol-specific biases.
A second aspect of the cooperation mechanism may collect summarized
network characteristics from other DLAs, and can answer queries as
to whether two DLAs are similar from the point of view of their
network traffic. In a first embodiment, a DLA itself may perform
such functions and may only answer queries as to whether another
DLA is similar to the DLA. In a second embodiment, the SCA or other
external device may answer any query regarding similarities. The
answer to the query is provided under the form of a similarity
score, in one embodiment.
The cooperation mechanism allows the DLA or SCA to determine
whether models from one DLA are relevant for use at another DLA.
The network characteristics may include summary statistics about
the hosts seen by the DLA, about the volume of traffic since at
different scales by the DLA, or details about applications seen by
the DLA (e.g., either application-classification levels metrics, or
port and IP protocol-based metrics). In another embodiment, a
statistical clustering model can be learned from the data provided
by the various DLAs to the SCA. Once properly trained, the model
can then be pushed to the DLAs, which may then provide the model ID
corresponding to the traffic class used for each model. The SCA may
then be able to determine which models may then be shared between
DLAs.
In various embodiments, the devices may exchange unicast or
multicast messages, AgentSimilarQuery( ) and AgentSimilarReply( ),
to query similarities. For example, as shown in FIG. 9B, DLA 400
may send an AgentSimilarQuery( ) message 904 to SCA 502, to
identify any other DLAs with similar models. In response, as shown
in FIG. 9C, SCA 502 may indicate that DLA 400b has one or more
similar models to those of DLA 400.
A third component of the cooperation mechanism may be a decision
system that monitors models deemed uncertain by the above
components and decides whether other DLAs have compatible models to
complement the incomplete model of the local DLA, and requests
these models.
The third component of the cooperation mechanism can decide not to
requests models, based on a cost function involving the output of
the first component, the similarity with other agents, but also the
immediate and historical bandwidth impact. In particular, bandwidth
caps can be fixed in order to avoid exchanging too many models at
agents that mostly have uncertain models. In addition, the cost
function may also be based in part on how often traffic
corresponding to the model has been seen in the past. In
particular, a very uncertain model for an application for which
only a couple packets have been seen may not trigger a request to
other agents to learn more. In some embodiments, unicast or
multicast messages, AgentModelRequest( ) and AgentModelResponse( ),
may be exchanged to query the models. For example, as shown in FIG.
9D, after DLA 400 determines that DLA 400b hosts a similar model,
DLA 400 may send an AgentModelRequest( ) message 908 to DLA 400b.
In response, DLA 400b may return the requested model to DLA 400 via
an AgentModelResponse( ) message 910.
In one embodiment gathering models is performed by a central
controller (e.g., an SCA, etc.). In another embodiment, upon
determining which DLA host models related to traffic that have
similar characteristics, the actual exchange of models may be
performed using a distributed approach, with direct communication
between DLAs. Such an approach allows for more efficient bandwidth
usage in the network and higher scalability.
FIG. 10 illustrates an example simplified procedure for using
anomaly detection models, in accordance with various embodiments
herein. Procedure 1000 may start at step 1005 and continues on to
step 1010 where, as described in greater detail above, a device in
a network may maintain a plurality of anomaly detection models. In
some cases, these models may be statistical models that analyze
different sets of aggregated traffic data from the network. For
example, one model may be based on both the source IP address of
the traffic, as well as traffic associated with a particular
application. In contrast, another model may be based on the
destination MAC address of the traffic for the particular
application.
At step 1015, as detailed above, the device may determine a measure
of confidence in one of the plurality of models. As would be
appreciated, the device may calculate any number of confidence
measures for a particular model. For example, the device may
calculate a statistical confidence value for a given statistical
model.
At step 1020, the device may dynamically replace the particular
model with another model, as described in greater detail above. In
various embodiments, the replacement model may have a different
capacity than that of the model being replaced. For example, if the
first model has a low confidence measure, the device may replace
the first model with a lower capacity model. Similarly, if the
first model has a high confidence measure, the device may opt to
replace this model with another model that has a higher
capacity.
At step 1025, as detailed above, the device may provide an anomaly
event notification to one or more other devices (e.g., an SCA).
Such a notification may be based on a combination of model outputs
from, e.g., the replacement model from step 1020 and one or more of
the other local models maintained by the device. In some cases, the
device may select only those models with the highest n-number of
confidence scores for combination. In another embodiment, the
device may base the selection on its own available resources, to
dynamically adjust the number of models allowed to contribute to
the notification. Procedure 1000 then ends at step 1030.
It should be noted that while certain steps within procedure 1000
may be optional as described above, the steps shown in FIG. 10 are
merely examples for illustration, and certain other steps may be
included or excluded as desired. Further, while a particular order
of the steps is shown, this ordering is merely illustrative, and
any suitable arrangement of the steps may be utilized without
departing from the scope of the embodiments herein.
The techniques described herein, therefore, provide a multi-scale
approach allows the system to raise very strong anomalies with
large confidence, as they may correspond to higher levels of
aggregation, but also more subtle anomalies with a lower
confidence. Further aspects of the techniques herein allows for the
selection of relevant events to raise anomalies using competition
buckets for statistical models. This allows the system to avoid
biases due to anomalies with large or low flow volume or durations,
as well as to avoid using an excessive amount of memory, which
would be especially troublesome on embedded deployments. Finally,
cooperation techniques are introduced herein that allow proper
predictions to be made in situations where some specific models
have only low amounts of samples due to the location of the agent
in the network, thus misrepresenting what can be observed over the
full network. This potentially leads to false positives, wherein a
behavior is scored as highly anomalous by the agent, while other
agents would not agree.
While there have been shown and described illustrative embodiments
that provide for anomaly detection, it is to be understood that
various other adaptations and modifications may be made within the
spirit and scope of the embodiments herein. For example, while
certain embodiments are described herein with respect to using
certain models for purposes of anomaly detection, the models are
not limited as such and may be used for other functions, in other
embodiments. In addition, while certain protocols are shown, such
as BGP, other suitable protocols may be used, accordingly.
The foregoing description has been directed to specific
embodiments. It will be apparent, however, that other variations
and modifications may be made to the described embodiments, with
the attainment of some or all of their advantages. For instance, it
is expressly contemplated that the components and/or elements
described herein can be implemented as software being stored on a
tangible (non-transitory) computer-readable medium (e.g.,
disks/CDs/RAM/EEPROM/etc.) having program instructions executing on
a computer, hardware, firmware, or a combination thereof.
Accordingly this description is to be taken only by way of example
and not to otherwise limit the scope of the embodiments herein.
Therefore, it is the object of the appended claims to cover all
such variations and modifications as come within the true spirit
and scope of the embodiments herein.
* * * * *
References